shumengya/sproutclaw
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add model thinking level metadata

Browse Source 
closes #3208
This commit is contained in:
Mario Zechner
2026-05-02 01:20:01 +02:00
parent 73b7b2c564
commit 80f06d3636
28 changed files with 527 additions and 243 deletions
Download Patch File Download Diff File Expand all files Collapse all files

50
packages/ai/src/providers/openai-completions.ts
View File

@@ -11,7 +11,7 @@ import type {
ChatCompletionToolMessageParam,
} from "openai/resources/chat/completions.js";
import { getEnvApiKey } from "../env-api-keys.js";
import { calculateCost, supportsXhigh } from "../models.js";
import { calculateCost, clampThinkingLevel } from "../models.js";
import type {
AssistantMessage,
CacheRetention,
@@ -36,7 +36,7 @@ import { parseStreamingJson } from "../utils/json-parse.js";
import { sanitizeSurrogates } from "../utils/sanitize-unicode.js";
import { isCloudflareProvider, resolveCloudflareBaseUrl } from "./cloudflare.js";
import { buildCopilotDynamicHeaders, hasCopilotVisionInput } from "./github-copilot-headers.js";
import { buildBaseOptions, clampReasoning } from "./simple-options.js";
import { buildBaseOptions } from "./simple-options.js";
import { transformMessages } from "./transform-messages.js";
/**
@@ -410,7 +410,8 @@ export const streamSimpleOpenAICompletions: StreamFunction<"openai-completions",
}
const base = buildBaseOptions(model, options, apiKey);
const reasoningEffort = supportsXhigh(model) ? options?.reasoning : clampReasoning(options?.reasoning);
const clampedReasoning = options?.reasoning ? clampThinkingLevel(model, options.reasoning) : undefined;
const reasoningEffort = clampedReasoning === "off" ? undefined : clampedReasoning;
const toolChoice = (options as OpenAICompletionsOptions | undefined)?.toolChoice;
return streamOpenAICompletions(model, context, {
@@ -547,21 +548,27 @@ function buildParams(
} else if (compat.thinkingFormat === "deepseek" && model.reasoning) {
(params as any).thinking = { type: options?.reasoningEffort ? "enabled" : "disabled" };
if (options?.reasoningEffort) {
(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);
(params as any).reasoning_effort =
model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;
}
} else if (compat.thinkingFormat === "openrouter" && model.reasoning) {
// OpenRouter normalizes reasoning across providers via a nested reasoning object.
const openRouterParams = params as typeof params & { reasoning?: { effort?: string } };
if (options?.reasoningEffort) {
openRouterParams.reasoning = {
effort: mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap),
effort: model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort,
};
} else {
openRouterParams.reasoning = { effort: "none" };
} else if (model.thinkingLevelMap?.off !== null) {
openRouterParams.reasoning = { effort: model.thinkingLevelMap?.off ?? "none" };
}
} else if (options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {
// OpenAI-style reasoning_effort
(params as any).reasoning_effort = mapReasoningEffort(options.reasoningEffort, compat.reasoningEffortMap);
(params as any).reasoning_effort = model.thinkingLevelMap?.[options.reasoningEffort] ?? options.reasoningEffort;
} else if (!options?.reasoningEffort && model.reasoning && compat.supportsReasoningEffort) {
const offValue = model.thinkingLevelMap?.off;
if (typeof offValue === "string") {
(params as any).reasoning_effort = offValue;
}
}
// OpenRouter provider routing preferences
@@ -583,13 +590,6 @@ function buildParams(
return params;
}
function mapReasoningEffort(
effort: NonNullable<OpenAICompletionsOptions["reasoningEffort"]>,
reasoningEffortMap: Partial<Record<NonNullable<OpenAICompletionsOptions["reasoningEffort"]>, string>>,
): string {
return reasoningEffortMap[effort] ?? effort;
}
function getCompatCacheControl(
compat: ResolvedOpenAICompletionsCompat,
cacheRetention: CacheRetention,
@@ -1057,32 +1057,13 @@ function detectCompat(model: Model<"openai-completions">): ResolvedOpenAIComplet
const useMaxTokens = baseUrl.includes("chutes.ai") || isMoonshot || isCloudflareAiGateway;
const isGrok = provider === "xai" || baseUrl.includes("api.x.ai");
const isGroq = provider === "groq" || baseUrl.includes("groq.com");
const isDeepSeek = provider === "deepseek" || baseUrl.includes("deepseek.com");
const cacheControlFormat = provider === "openrouter" && model.id.startsWith("anthropic/") ? "anthropic" : undefined;
const reasoningEffortMap = isDeepSeek
? {
minimal: "high",
low: "high",
medium: "high",
high: "high",
xhigh: "max",
}
: isGroq && model.id === "qwen/qwen3-32b"
? {
minimal: "default",
low: "default",
medium: "default",
high: "default",
xhigh: "default",
}
: {};
return {
supportsStore: !isNonStandard,
supportsDeveloperRole: !isNonStandard,
supportsReasoningEffort: !isGrok && !isZai && !isMoonshot && !isCloudflareAiGateway,
reasoningEffortMap,
supportsUsageInStreaming: true,
maxTokensField: useMaxTokens ? "max_tokens" : "max_completion_tokens",
requiresToolResultName: false,
@@ -1118,7 +1099,6 @@ function getCompat(model: Model<"openai-completions">): ResolvedOpenAICompletion
supportsStore: model.compat.supportsStore ?? detected.supportsStore,
supportsDeveloperRole: model.compat.supportsDeveloperRole ?? detected.supportsDeveloperRole,
supportsReasoningEffort: model.compat.supportsReasoningEffort ?? detected.supportsReasoningEffort,
reasoningEffortMap: model.compat.reasoningEffortMap ?? detected.reasoningEffortMap,
supportsUsageInStreaming: model.compat.supportsUsageInStreaming ?? detected.supportsUsageInStreaming,
maxTokensField: model.compat.maxTokensField ?? detected.maxTokensField,
requiresToolResultName: model.compat.requiresToolResultName ?? detected.requiresToolResultName,